Update README.md
Browse files
README.md
CHANGED
|
@@ -39,6 +39,8 @@ An advanced deep learning model for automatically tagging anime/manga illustrati
|
|
| 39 |
- **Streamlit web interface**: User-friendly UI for uploading and analyzing images.
|
| 40 |
- **Adjustable threshold profiles**: Overall, Weighted, Category-specific, High Precision, and High Recall profiles.
|
| 41 |
- **Fine-grained control**: Per-category threshold adjustments for precision-recall tradeoffs.
|
|
|
|
|
|
|
| 42 |
|
| 43 |
## Performance Notes:
|
| 44 |
|
|
@@ -554,6 +556,7 @@ The monitor provides invaluable insights into how the two-stage prediction model
|
|
| 554 |
## Acknowledgments:
|
| 555 |
|
| 556 |
- Claude Sonnet 3.5 and 3.7 for being incredibly helpful with the brainstorming and coding.
|
|
|
|
| 557 |
- [Danbooru](https://danbooru.donmai.us/) for the incredible dataset of tagged anime images.
|
| 558 |
- [p1atdev](https://huggingface.co/p1atdev) for the processed Danbooru 2024 dataset.
|
| 559 |
- Microsoft for DeepSpeed, which made training possible on consumer hardware.
|
|
|
|
| 39 |
- **Streamlit web interface**: User-friendly UI for uploading and analyzing images.
|
| 40 |
- **Adjustable threshold profiles**: Overall, Weighted, Category-specific, High Precision, and High Recall profiles.
|
| 41 |
- **Fine-grained control**: Per-category threshold adjustments for precision-recall tradeoffs.
|
| 42 |
+
- **Safetensors and ONNX**: Original pickle files available in /models.
|
| 43 |
+
- **EfficientNetV2-L Backbone**": Backbone performance greatly improved by the refining embedding layer.
|
| 44 |
|
| 45 |
## Performance Notes:
|
| 46 |
|
|
|
|
| 556 |
## Acknowledgments:
|
| 557 |
|
| 558 |
- Claude Sonnet 3.5 and 3.7 for being incredibly helpful with the brainstorming and coding.
|
| 559 |
+
- [EfficientNetV2](https://arxiv.org/abs/2104.00298) for the backbone
|
| 560 |
- [Danbooru](https://danbooru.donmai.us/) for the incredible dataset of tagged anime images.
|
| 561 |
- [p1atdev](https://huggingface.co/p1atdev) for the processed Danbooru 2024 dataset.
|
| 562 |
- Microsoft for DeepSpeed, which made training possible on consumer hardware.
|